ERP-based Speller with a New Paradigm
Jin-Hun Sohn, Mi-Sook Park, Hye-Ryeon Yang, Young-Ji Eum and Jin-Sup Eom
Department of Psychology/Brain Research Institute, Chungnam National University, Daejeon, South Korea
Keywords: P300 Speller, Brain-computer Interface, ERP.
Abstract: In most implementation of an ERP-based speller, standard row-column paradigm (RCP) was used.
However, RCP is susceptible to adjacency-distraction errors because items in the same row or column of the
target flash at the time of a half when the target item flashes. The adjacency-distraction errors could be
reduced if the number of flanking items that flash with the target is diminished. This study presents a novel
P300-based stimulus presentation called row-column-diagonal paradigm (RCDP) where characters on the
main diagonal and the anti-diagonal in the matrix flash in addition to characters on the row and columns. In
RCDP, items in the same row, column, main diagonal, and anti-diagonal of the target flashes at the time of a
quarter when the target item flashes. Using a 6×6 matrix of alphanumeric characters and keyboard
commands, ten college students used RCP and RCDP. Stepwise linear discriminant analysis (SWLDA) for
the EEG signals recorded in calibration phases was used to calculate discrimininant function. By applying the
discrimininant function to electroencephalography (EEG) signal recorded in the test phase, the probability
whether the item was the target or not was evaluated. Average accuracy was 76.6% in RCP while 84.0% in
RCDP. With RCP, most errors were occurred in the same row or column of the target; on the other hand,
with RCDP in the same row, column, main diagonal, or anti-diagonal of the target. These findings indicate
how RCDP reduces adjacency-distraction errors and might be able to contribute to develop more advanced
stimulus presentation paradigm.
1 INTRODUCTION
Although it had been almost impossible for the
paralyzed patients to communicate with others and
to control any devices, Farwell and Donchin (1988)
introduced an innovative invention for them to
communicate by using electroencephalography. The
invention was the first speller paradigm that is to
write characters in a computer by using event-related
potential (ERP). In this paradigm, a computer
presents a 6 × 6 matrix of letters on screen as shown
in Fig 1 (A) and groups of matrix items flash
randomly. Twenty trials should be performed in this
paradigm to spell one character where one sequence
consisted of the six rows in the matrix flash in order
followed by the six columns in the matrix flash.
There participants are asked to attend to the item they
wish to select or count the number of times the item
flashes.
The amplitude of P300, one of the ERP
components increases when a participant attends to
the item. If a computer calculate the P300 amplitude
of users when a item flash, the computer is able to
identify the attended item as the intersection of the
row and column that elicited the largest P300. Due to
the P300 response’ s relatively low signal-to-noise
ratio, each item must be flashed multiple times and
the results averaged. The number of times the item
flashes is positively correlated with average accuracy
(Donchin et al., 2000; Lenhardt et al., 2008). The
more each item flashes, the less ERP noise occur and
the higher accuracy become. On the other hand, the
more each item flashes, the longer time for
presentation becomes for a participant to spell a
character.
The novel paradigm developed by Farwell and
Donchin(1988) is called the row column paradigm,
or RCP since items are grouped for flashing as rows
and columns. However, the RCP remains subject to
errors and, importantly, these errors appear to have
one primary cause (Fazel-Rezai, 2007; Fazel-Rezai
et al., 2012; Townsend et al., 2010). Errors typically
occur with the greatest frequency in locations
adjacent to the attended item (i.e., the target item)
and almost always in the same row or column. This
inherent RCP error occurs because adjacent item
flashes, it attracts the participants’ attention
341
Sohn J., Park M., Yang H., Eum Y. and Eom J..
ERP-based Speller with a New Paradigm.
DOI: 10.5220/0004834203410346
In Proceedings of the International Conference on Physiological Computing Systems (PhyCS-2014), pages 341-346
ISBN: 978-989-758-006-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
(Townsend et al., 2010). We refer to these relatively
systematic errors as adjacency-distraction errors”.
There is additional cause in RCP for the occurrence
of adjacency-distraction errors because items in the
same row or column of the target flash at the time of
a half when the target item flashes. It consequently
results in ERP response.
This study is to investigate a novel stimulus
presentation paradigm that can control adjacency-
distraction errors by reducing the number of flanking
items that flash with the target. In the new paradigm
as shown in Figure 1 (B), the target item on the main
diagonal in the matrix flash in order followed by the
anti-diagonal in the matrix. The paradigm is called
row-column-diagonal paradigm (RCDP). With RCP,
items are grouped for flashing as 6 rows and 6
columns in one trial. Items are grouped for flashing
as 6 rows, 6 columns, 6 main diagonal, and 6 anti-
diagonal in one trial with RCDP.
In RCP, items in the same row or column of the
target flashes at the time of a half when the target
item flashes while in RCDP, items in the same row,
column, main diagonal, and anti-diagonal of the
target flashes at the time of a quarter when the target
item flashes. In this study, we hypothesized that the
RCDP will produce superior performance as
compared to the RCP because it diminishes the
adjacency-distraction errors to which the RCP is
prone.
2 METHOD
2.1 Participants
10 college students (5 males) participated in this
experiment and their mean age was 24.4 years
(range 22-28). They had no experience to participate
in ERP-based speller experiment before. They had
normal or corrected-to-normal vision.
2.2 Equipment
Each participant sat in a comfortable chair
approximately 60 cm from a 19 inch LCD monitor
with a 1280×1024 resolution that displayed the 6×6
matrix. The width of each character included in the
matrix was 1.1cm and height was 1.3cm and the
space between characters was 5cm on the right and
left, 3cm on the top and bottom. According to
Krusienski et al., (2008) results, EEG activity was
recorded from Fz, Cz, Pz, Oz, P3, P4, PO7, and
PO8. Linked electrodes attached to the mastoids
served as reference and the ground electrode was
placed at the forehead. The signals were amplified
using a Grass Model 12 Neurodata Acquisition
System (Grass Instruments, Quincy, MA, USA)
(high-pass and low-pass filters 0.3 and 30Hz,
respectively) with 20000 amplification rate. EEG
was recorded by bio-amplifier MP150 (BioPac
Systems Inc., Santa Barbara, CA, USA) and the
signals were saved at the sampling rate of 200Hz.
Recording programs for stimulus presentation and
EEG activity was created via visual C++ v6.
2.3 Experimental Procedure
Each participant completed two experimental
sessions. Sessions began with the RCP session and
the RCDP session followed. Each session consisted
of a calibration phase and an test phase. The first
phase was a calibration phase to generate
discrimininant function for identifying target item.
The second phase was a test phase for detecting the
target item by applying the discrimininant function.
Total 18 items were used in a calibration phase
and 25 items consisted of rows having 5 characters
and 1 number in a test phase. In RCP, one row or
column from the 6×6 matrix flashes once at a time in
a random order. The participant’s task was to attend
to (or count) the number of times the item in a row or
column flashed.
When spelling a character, one trial is defined
where 6 rows and 6 columns flashes all once at a
time and total 6 trials were repeated. In RCDP, 6
rows, 6 columns, 6 main diagonal, and 6 anti-
diagonal flash in a random order. In RCDP, one trial
is defined where, 6 rows, 6 columns, 6 main
diagonal, and 6 anti-diagonal flashes all once at a
time and total 3 trials were repeated. In both RCP
and RCDP, each set of items flashed for 100ms
followed by a 25ms inter-stimulus interval. Sessions
were counterbalanced to minimize the effect of the
order. After completing both sessions, participants
were asked to rate about how difficult it was when
performing each type of paradigm on 7-point Likert
scale, with 1=Very difficult and 7=Very easy.
2.4 Classification
Stepwise linear discriminant analysis (SWLDA) for
the EEG signals recorded in calibration phases was
used to calculate discrimininant function. The
probability whether the item was the target or not
was calculated by applying the discrimininant
function to EEG signal recorded in the test phase.
PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems
342
(A) (B)
Figure 1: (A) Row-column paradigm. (B) Row-column and diagonal paradigm.
SWLDA was conducted as follows. When
spelling a character, rows and columns flash 72 times
(In RCDP, rows, columns, main diagonal, and anti-
diagonal) flashed. During this period, EEG activity
was recorded from 8 scalp locations. After a set of
items start to flash, 750ms data segment was created.
Seventy-two segments were made while one
character flashes. One segment recorded from an
electrode consists of 150 values (750ms × 200Hz).
These segments are distinguished in case where
the target item is included from where the target
item is not included. As a result, a matrix having 72
× 18 columns and 150 × 8 rows was created, and
discriminant function that can distinguish target item
from non-target item was calculated by conducting
SWLDA
In test phase, an participant spell a character ERP
was calculated for the each 36 characters on the 36
× 1200 matrix by averaging EEG segment recorded
while each item flickers.
The probabilities whether the column was the
target or not was calculated by applying the
discrimininant function for the 36 columns. The item
that shows the highest probability was selected as
target characters.
2.5 Transfer Rate
A possible method for evaluating the speller function
is the amount of information that is conveyed per
time unit, also known as data transfer rate or bit rate.
The written symbol rate (WSR) can be determined
by first computing the bits (B) per trial and then the
symbol rate (SR; see below the fomula 1)
2
2
1
2
1
1
(1)
N is the number of possible targets and P is the
probability that the target is accurately classified.
Then from equation (1) the symbol rate is determined
as,
2
(2)
If T is the trial duration in minutes, the WSR can be
determined as follows:
21
0
0.5
0.5
(3)
ERP-basedSpellerwithaNewParadigm
343
3 RESULTS
Table 1 shows accuracy, B, and WSR. Average
accuracy was 72.8% in RCP while 83.2% in RCDP.
The RCDP was significantly more correct in P300
speller performance than the RCP(t(9)=2.57, p<.05).
B was significantly lower in RCP as of 20.3% than
in RCDP as of 25.3%.(t(9)=2.58, p<.05). The WSR
in RCDP as of 3.5 was significantly higher in RCP
as of 1.8 (t(9)=2.35, p<.05).
In the individual data, eight out of ten
participants showed higher accuracy in RCDP than
in RCP, one showed the same accuracy in both
paradigms, and the last one showed higher accuracy
in RCP than in RCDP. Figure 2 represent errors
occurring in each cell relative to the target location.
Sixty-eight errors were occurred in RCP where 62
out of them (91%) occurred in the same row or
column of the target. Forty-two errors were occurred
in RCDP where 35 out of them (83%) occurred in
the same row, column, main diagonal, or anti-
diagonal of the target.
When the participants rated about how difficult it
was when using each type of paradigm, their mean
score was 5.1 in RCP and 4.9 in RCDP. There was
no significantly difference (t(9)=.557, n.s.).
4 CONCLUSIONS
ERP-based speller by using RCP is susceptible to
adjacency-distraction errors. This study presents a
alternative P300-based stimulus presentation–RCDP
to avoid the errors and examine average accuracy.
The results showed that, as we hypothesized, the
RCDP was more correct in P300 speller
performance than the RCDP.
Average accuracy was 72.8% in RCP while
83.2% in RCDP. Guger et al. (2009) reported 91%
of average accuracy in RCP based on the findings
from offline experiment among 81 healthy subjects.
It was higher than the result of this study, and the
reason might be the number of trials they used, 20
trials for a character. If the number of trials in RCDP
is calculated to that of RCP, RCDP uses 6 trials for a
character. When the number of trials increases, the
accuracy rate becomes greater (Furdea, 2009). Thus,
if the same number of trials is used for RCDP, it will
result higher accuracy than that of RCP.
Townsend et al. (2010) compared accuracy rate
of RCP to that of checkerboard paradigm (CBP), an
alternative paradigm which they designed. When 3
or 5 trials were used, average accuracy was 77.3% in
RCP while 91.5% in CBP. They found that the CBP
yielded higher accuracy than that of the RCDP.
There are two main reasons for higher accuracy of
CBP than the RCDP. Firstly, this study used an 6×6
matrix, while Townsend et al. (2010) used 8×9
matrix. The 6×6 matrix needs to be flashed 6 times
to intensify each character, whereas the 8×9 needs to
be flashed 12 times. Thus, inter-target interval (ITI)
of CBP becomes longer by using 8×9 matrix than
that of RCDP that uses 6×6 matrix. When the ITI
becomes longer than 5 seconds, ITI does not affect
P300 amplitude (Rasmusson and Allen 1994).
However, when the ITI is shorter than 1 second, ITI
affect P300 amplitude. P300 responses to the current
target elapse since the responses overlapped with
those of the previous target.
For this reason, in CBP where ITI is more than
one second, P300 amplitude becomes larger and
accuracy rate increases. Secondly, CBP controls
double-flash error effectively. Although RCDP is
efficient for controlling adjacency-distraction errors,
it still has double-flash problem.
Table 1: Values of accuracies and bit rates (bits/min) for the RCP and RCDP.
Participant
RCP RCDP
Accuracy Bit rate WSR Accuracy Bit rate WSR
1 68.0 17.5 0.1 96.0 31.5 5.5
2 92.0 29.0 4.6 100.0 34.5 6.7
3 48.0 10.0 0.0 52.0 11.4 0.0
4 80.0 22.8 2.2 88.0 26.8 3.7
5 64.0 15.9 0.0 64.0 15.9 0.0
6 92.0 29.0 4.6 88.0 26.8 3.7
7 68.0 17.5 0.1 92.0 29.0 4.6
8 56.0 12.8 0.0 88.0 26.8 3.7
9 100.0 34.5 6.7 100.0 34.5 6.7
10 60.0 14.3 0.0 64.0 15.9 0.0
Mean 72.8 20.3 1.8 83.2 25.3 3.5
PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems
344
(A) (B)
Figure 2: (A) Error (A) Error distributions for the RCP. (B) Error distributions for the RCDP. All target items have been
centered in each matrix; the number in the black cell is the number of correct selections for each paradigm. Numbers in
other cells represent the number of errors occurring in each cell relative to the target location.
One advantage of RCDP compared to CBP is
that it uses smaller matrix. By using smaller matrix,
RCDP needs less number of trials for spelling an
item compared to CBD which uses larger matrix.
The results of an error analysis from the RCP
were consistent with previous studies (Fazel-Rezai,
2007; Townsend et al., 2010). More than 90% of
errors were occurred in the same row or column of
the target. In RCP, items in the same row or column
of the target flash at a rate of 50% when the target
item flashes. On the other hand, in RCDP items in
the same row or column of the target flash at a rate
of 25% at most when the target item flashes.
Consequently, this novel paradigm reduced
adjacency-distraction errors by diminishing the
flickering frequency for adjacent items to the target
item when the target item flashes.
The performance accuracy is highly related to the
flickering frequency of the each character. To
sustain the stable accuracy, trials should be repeated
less in RCDP while the trials should be repeated
more in RCP. These findings suggest how RCDP
reduces adjacency-distraction errors and might be
able to contribute to develop better stimulus
presentation paradigm.
ACKNOWLEDGEMENTS
This research was supported by the Converging
Research Center Program through the Ministry of
Science, ICT and Future Planning, Korea
(2013K000332).
REFERENCES
Donchin, E., Spencer, K. M., & Wijesinghe, R., 2000, The
mental prosthesis: Assessing the speed of a P300-
based brain-computer interface, IEEE Transactions on
Rehanilitation Engineering, 8, 174-179.
Farwell, L. A., & DOnchin. E., 1988. Talking off the top
of your head: Toward a mental prosthesis utilizing
event-related brain potentials, Electroencephalo-
graphy and Clinical Neurophysiology, 70, 510-523.
Fazel-Rezai, R., 2007. Human error in P300 speller
paradigm for brain-computer interface, In Engineering
in Medicine and Biology Society, EMBS 2007, 29th
Annual International Conference of the IEEE(pp.
2516-2519).
Fazel-Rezai, R., Allison, B. Z., Guger, C., Sellers, E. W.,
Kleih, S. C., & Kubler, A., 2012. P300 brain computer
interface: Currnet challenges and emerging trends,
Frontiers in Neuroengineering, 5, 1-14.
Furdea, A., Halder, S., Krusienski, D. J., Bross, D.,
NijBoer, F., Birbaumer, N., & Kubler, A., 2009. An
suditory oddball (P300) spelling system for brain-
computer interfaces, Psychophysiology, 46, 617-625.
Guger, C., Daban, S., Sellers, E., Holzner, C., Krausz, G.,
Carabalona, R., Gramatica, F., & Edlinger, G., 2009.
How many people are able to control a P300-based
brain-computer interface (BCI)?, Neuroscience Letters,
463, 94-98.
ERP-basedSpellerwithaNewParadigm
345
Lenhardt, A., Kaper, M., & Ritter, H. J., 2008. An adapted
P300-based online brain-computer interfafce, IEEE
Transactions on Neural System and Rehabilitation
Engineering, 16, 121-130.
Townsend, G., Lapallo, B. K., Boulav, V. B., Krusienski,
D. J., Frye, G. E., Hauser, C. K., Schwartz, N. E.,
Voughan, T. M., Wolpaw, J. R., & Sellers, E. W.,
2010. A novel P300-based brain-computer interface
stimuluis presentation paradigm: Moving beyond rows
and columns, Clinical Neurophysiology, 131, 1109-
1120.
Rasmusson, D. X., & Allen, J. D., 1994. Intertarget
interval does not affect P300 within oddball series,
International Journal of Psychophysiology, 17, 57-63.
Smith, J., 1998. The book, The publishing company.
London, 2
nd
edition.
Krusienske, D. J., Sellers, E. W., McFarland, D. J.,
Vaughan, T. M., & Wolpaw, J. R., 2008. Toward
enhanced P300 speller performance, Journal of
Neuroscience Methods, 167, 15-21.
PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems
346
